The paper introduces Complex KDA (CKDA), an enhanced Kimi Delta Attention variant for linear RNNs that extends parameter ranges (gates in [-1,1], delta-rule coefficients in [0,2]) so the mechanism can realize 2D rotations while remaining stable and efficient. The authors prove every orthogonal diagonal-plus-rank-one matrix corresponds to a CKDA transition matrix, giving stronger state-tracking over finite groups and improved length-extrapolation and language-modeling results.
